On September 29, 2026, the US Court of Appeals for the Third Circuit issued its long-awaited decision in the lawsuit of Thomson Reuters v ROSS Intelligence. This case was filed in 2020 and was first mentioned in the Scholarly Kitchen in March 2023. As of today, it is the only Circuit Court decision in the US to address training. It was a big win for copyright owners.
Background
The case involved the surreptitious copying of the entire Westlaw legal database (after having been denied a license) in order to create a competing “AI legal search engine.” The original defendant went out of business but continues (somehow) to litigate. After Thomson Reuters (TR) won summary judgement, ROSS Intelligence (ROSS) appealed on two issues; copyrightability of the copied works and fair use.

Copyrightability
Expression is protected by copyright. Facts are not. Nor are US Government Works such as statutes and cases.
At issue here was Westlaw’s headnotes, containing summaries of public domain cases (vaguely similar to abstracts in journals). While text of headnotes is largely driven by the facts of the cases they summarize, the court took note of TR’s editorial practices and held:
All 2,243 independently created headnotes at issue in this appeal have ‘some creative spark.’ For each headnote, Thomson Reuters’s editors made creative decisions about which points of law were important enough to include in a headnote and how to word those points of law consistent with Westlaw’s drafting criteria. Each headnote had to convey enough factual background and reasoning to stand on its own….”
This ruling is useful and potentially relevant to journal and standards publishers. Courts will at times afford less protection given the “factual” nature of the publications while ignoring the difference between expression and facts. This case should help publishers win that argument.
Infringement and Fair Use
Having established that the headnotes were entitled to protection, the court looked at whether the affirmative defense of fair use applied. Fair use is one of the most mythologized concepts in copyright law. To myth bust, please keep in mind fair use is not a “right” (it is an affirmative defense; essentially an excuse for infringement and the burden of proving the excuse is on the infringer), does not simply attach to new technologies (technologies such as AI are not uses; uses are the things you do with the AI), and transformative uses are not fair use per se (courts will look to it, but it is just part of a larger analysis).
Every fair use case involves a fact-dependent analysis of four factors set forth in the Copyright Act. This is why rulings can seem so inconsistent. And while copyright lawyers always say that fair use involves an analysis of four factors, the action is really in factors 1 (nature of the use, such as commercial/noncommercial and transformative/non-transformative) and 4 (market harm).
Minimally Transformative
As the use was clearly commercial, ROSS tried to escape liability by arguing that AI is transformative. The court found that the Defendant’s use was “minimally transformative, at best,” reasoning:
Thomson Reuters uses its headnotes to help legal researchers find and understand judicial opinions containing points of law related to their searches. ROSS uses Thomson Reuters’s headnotes to create a legal-research platform to help users find judicial opinion passages responsive to their searches. True, ROSS did not use the headnotes in precisely the same fashion as Thomson Reuters. ROSS took an intermediate step of using the headnotes to train an AI program, which arguably presents a slight degree of difference in use. But the undisputed evidence demonstrates that ROSS used the headnotes to train an AI program for the benefit of its legal-research platform. So both Thomson Reuters and ROSS use the headnotes to create and optimize a legal-research platform that helps users find responsive legal material.
Let’s pause on this text. Big Tech loves to promise that AI will cure cancer — at least when it is not asserting that AI will kill us all. Let’s rerun the quote with some terms swapped:
[Oncology Publisher] uses its [journals] to help [oncology] researchers find and understand [science] containing points [] related to their searches. [Tech Firm] uses [Oncology Publisher]’s [journals] to create an [oncology] research platform to help users find [research] responsive to their searches. True, [Tech Firm] did not use the [journals] in precisely the same fashion as [Oncology Publisher]. [Tech Firm] took an intermediate step of using the [journals] to train an AI program, which arguably presents a slight degree of difference in use. But the undisputed evidence demonstrates that [Tech Firm] used the [journals] to train an AI program for the benefit of its [] research platform. So both [Oncology Publisher] and [Tech Firm] use the [journals] to create and optimize a [] research platform that helps users find responsive [oncology] material.
In other words, the Third Circuit’s opinion supports the position that the more the AI is marketed as solving research challenges, the less transformative it becomes with respect to STM content.
Market Harm
In the end, market harm is generally regarded as the most important fair use factor. Here the court looked both at the original market for legal research and the potential market for content licensing.
Given that ROSS was trying to directly compete with TR in the market for legal research, the court easily concluded there was harm in the original market. However, the court went further:
Here, the evidence shows that the market for licensing headnotes as text to train AI is rapidly developing. Additionally, Thomson Reuters presented evidence that it is using its headnotes as training data for its own AI search products. ROSS offers no evidence to disprove this. That Thomson Reuters did not license its headnotes to others does not disprove that a market exists to do so. Additionally, ROSS’s copying harmed Thomson Reuters in this market. By copying the headnotes for use as training data without Thomson Reuters’s authorization, ROSS usurped Thomson Reuters’s opportunity to enter that derivative market and license its headnotes for that purpose. The market harm caused by ROSS’s use weighs against fair use. (emphasis added)
Remember that this case dates back to 2020. Back then, TR was not licensing content for AI. It still won on its right to enter into the market. Contrast that with the journals sector. CCC launched its collective license for text and data mining in 2015. We added AI rights to our corporate license in 2024. Publishers have likewise been directly licensing AI/TDM (text- and data-mining) rights for at least that long.
There is no better evidence of a potential licensing market than an actual one and here, our industry is factually strong.
“But the ROSS Case is not Generative AI”
Ross went out of business in 2020, yet somehow money magically appeared to hire three of America’s most prestigious (and expensive) law firms to file a petition for rehearing. Clearly Silicon Valley believes it has a lot at stake. That said, their spin doctors like to pretend this case was irrelevant because as the court noted, this case did not involve generative AI.
True, but much AI is not generative AI. The AI in my search engine yields results relevant to my query; it does not generate a term paper. Like ROSS’s product, it provides me with the answer I seek from the literature. For AI firms to support trillion-dollar valuations, they will need business customers. As noted in a recent study entitled The Velocity of Content in the Age of AI 2026 Copyrighted Content Usage Trends, business users often use AI to answer questions, which is exactly what this case involved.
Moreover, generative AI poses the risk of copies being made in outputs. If the case did involve generative AI, the market harm would have been increased by the unlimited number of infringing outputs based on the headnotes, increasing the harm. Lawyers gotta try, but life is not that simple.
Applying Other Cases
Having criticized those who declare AI training to categorically be fair use, I will be the last to argue it will never be fair use. Was the infringed content a Reddit post, an out-of-print book, or a journal article available for TDM rights under license? Does the AI use compete with an actual licensing market, a potential licensing market, no licensing market? All of these factors matter.
Those who state that “everything is fair use” like to cite the Google Books case. While that case held scanning books for the purposes set forth (which were much more limited than what is involved in the 150+ other US cases on AI and copyright) was deemed fair use, those people conveniently forget that Google “voluntary” ceased scanning journals during the pendency of the case? Did they do this because Google no longer wanted the journals, or because journal publishers were already licensing back files and thus it undercut the fair use argument?
I cannot answer that question. I will point out that the same court ruled against fair use two years later in a case involving similar practices with videos. Facts (as well as lawyers and judges) matter. Hence inconsistent rulings on AI training issued from the same District Court in California the same week in 2025.
When Will it End?
I have heard many discussions about when the application of fair use to AI training will be “settled law.” Fair use is part of the Copyright Act and as such is settled law. It will never, however, be applied in all AI training cases in one monolithic way or the other. Fair use is an analysis of four factors with many sub-factors. Unlike copies of content made for AI training, it is not binary.